Kernel fitting for speech detection and enhancement

A kernel fitting algorithm is proposed for speech denoising to improve the precision of voice activity detection (VAD) and the performance of speech enhancement, of some popular algorithms. In the algorithm, a noisy speech frame is filtered by kernel fitting, and then its power spectral density is estimated and weighted by a gain factor constructed from frame energy and zero-crossing rate, so that a speech signal is obviously discriminated from a nonspeech one. By incorporation of the VAD outputs and the noise effect into the kernel fitting process, a speech frame is enhanced with better performance than the spectra subtraction algorithm. Experiments are taken on a real life speech signal plus simulated noises, and the results show the potentiality of the proposed algorithms in speech detection and enhancement.